Why Feature Drift Is Harder Than Data Drift

Data drift = inputs change
Feature drift = the logic that generates the feature changes

Example:
Your “active user” feature used to be “clicked in last 7 days.”
Marketing redefines it to “clicked in last 3 days.”
Your model silently dies because the underlying concept changed.

Feature drift is more dangerous:
it happens inside your system, not in external data.

Production ML must version:
▪️feature definitions
▪️transformation logic
▪️data contracts

Otherwise the same model receives different features week to week.